Pith. sign in

REVIEW 5 major objections 5 minor 1 cited by

The paper claims that diffusion inverse solvers improve when inference-time search over candidate reconstructions is guided by side information such as a reference image, text caption, or MRI contrast—without retraining or reward gradients.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Injecting side information via inference-time particle search (GS/RFJS) improves diffusion-based inverse problem reconstructions across inpainting, super-resolution, deblurring, and MRI tasks in a training-free, plug-and-play way.

T0 review reviewed 2026-08-04 challenge →

load-bearing objection Solid training-free side-information search for diffusion solvers; the empirical scope is impressive, but the 'consistent improvement' claim needs error bars and a negative-control before it fully holds. the 5 major comments →

arxiv 2510.03352 v3 pith:55X6Q3SY submitted 2025-10-02 cs.CV cs.AIcs.LG

Inference-Time Search Using Side Information for Diffusion-Based Image Reconstruction

Mahdi Farahbakhsh , Vishnu Teja Kunde , Dileep Kalathil , Krishna Narayanan , Jean-Francois Chamberland This is my paper
classification cs.CV cs.AIcs.LG
keywords diffusion modelsinverse problemsside informationinference-time searchreward-guided resamplingimage reconstructiondeblurringMRI reconstruction
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper tries to establish that existing diffusion-based solvers for inverse problems can exploit arbitrary side information at inference time, with no retraining and no gradient backpropagation through the reward. It encodes side information as a reward, tilts the diffusion prior toward high-reward reconstructions, and searches over a small population of candidate particles using scheduled group resampling. Across box inpainting, super-resolution, several deblurring tasks, and accelerated MRI, the augmented solvers are reported to consistently beat the original solvers and also beat reward-gradient guidance. If correct, this gives a modular way to inject auxiliary knowledge into any existing diffusion reconstruction pipeline.

Core claim

The central discovery is that sampling from the side-information-conditioned posterior can be approximated without learning a conditional model. The paper models the conditional prior as p0|S(x0|s) ∝ p0(x0) exp(r(x0;s)/τ), where r is any reward function scoring compatibility between a reconstruction and the side information. Instead of taking gradients of this tilted score, which requires second-order derivatives through the denoiser, the paper evolves N candidate particles under the base solver and resamples them at scheduled steps according to reward weights. Greedy search resamples the entire group at fixed intervals; recursive fork-join search resamples hierarchically nested group sizes

What carries the argument

The central object is the reward-tilted conditional distribution p0|S(x0|s) ∝ p0(x0) exp(r(x0;s)/τ), with a pretrained reward function such as an identity-embedding distance for face images, a text-image alignment score for captions, or normalized mutual information for MRI contrasts. The second central mechanism is grouped resampling: at chosen reverse-diffusion times, candidate particles are ranked by reward, and within each group the highest-reward particle is replicated while others are discarded. Recursive fork-join search's hierarchical schedule of group sizes is what carries the exploration-exploitation balance; the paper's Proposition 1 provides the tilting justification, and its err

Load-bearing premise

The load-bearing premise is that the chosen reward function r(x0;s) orders candidate reconstructions by their true compatibility with the side information, so that resampling toward higher reward reliably pulls the sample toward the ground truth rather than toward side-information-specific artifacts.

What would settle it

Run the search module with a deliberately mismatched side-information reward—say, a reference face of a different identity or a text caption describing a different object—on a set of test images and compare against the base solver. If the side-information-guided reconstructions are not closer to the ground truth than the unguided ones, the core premise that reward resampling steers toward truth is falsified.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • The search module is a plug-in: an existing diffusion solver can be wrapped without modifying its denoising procedure, so gains transfer across solvers rather than requiring a new one.
  • Because the reward is used only to rank and resample candidates, non-differentiable and black-box side-information scores can be used, with no backpropagation through the diffusion network.
  • The same module handles image, text, and MRI-contrast side information, making the approach modality-agnostic in practice.
  • In severely ill-posed settings where the measurement alone leaves many plausible reconstructions, side-information search is claimed to reduce ambiguity and preserve identity or semantic content better than the base solver.
  • Reward-gradient guidance is shown to be brittle to gradient scale and to alter mostly fine details, whereas search can change coarse structure by switching candidate trajectories.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Since the method's behavior is entirely controlled by the reward, a natural test is to run it with deliberately mismatched side information—a reference image of a different identity, or a caption describing a different object. Nothing in the algorithm checks reward monotonicity, so the search would plausibly steer toward the side information rather than the ground truth; the paper does not report
  • The authors' own limitation section concedes that the exploration-exploitation schedule has no formal optimality guarantees; the recursive fork-join schedule and resampling interval are tuned heuristics. A testable extension is to derive schedules that adapt the resampling frequency to the degree of ill-posedness.
  • The nested group sizes in recursive fork-join search resemble a cooling schedule, with the temperature τ and base B acting as exploration controls. One could test whether annealing τ as the reverse process approaches t=0 improves the reconstruction-robustness tradeoff.
  • The same reward-tilting plus particle-search recipe may transfer to other inverse problems where a cheap scoring function exists—audio-to-image alignment, depth-guided photography, or text-guided medical imaging—provided the scoring function is monotone in true compatibility.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

5 major / 5 minor

Summary. This paper proposes a training-free, plug-and-play inference-time search module for diffusion-based inverse problem solvers. Side information (reference image, text, MRI contrast) is encoded as a reward, and the target conditional distribution is modeled as a reward-tilted prior (Eq. (2)). The authors derive a particle-filter-style update (Proposition 1) and instantiate two search strategies: Greedy Search (GS) and Recursive Fork-Join Search (RFJS). They evaluate on six inverse problems (box inpainting, super-resolution, motion/Gaussian/nonlinear/blind deblurring, MRI) with three base samplers (DPS, DAPS, MPGD) and report improvements over the base solvers, Best-of-N, and reward-gradient guidance on face, text, and MRI tasks.

Significance. The idea of leveraging side information at inference time without retraining is timely and practically appealing. If the improvements are robust, the framework could be dropped into existing diffusion pipelines with minimal modification. The paper's strengths include the breadth of experiments (multiple samplers, tasks, and side-information modalities), the clear modular formulation, and the release of code. However, the central empirical claim of 'consistently improves' is weakened by the absence of error bars/significance tests in the main tables, by the greedy implementation not matching the weighted-resampling theory, and by the untested reward-alignment assumption. The approach is a heuristic guided by a sensible but unvalidated model, so the theoretical contribution is limited.

major comments (5)
  1. [§5.2, Tables 1 and 2] The abstract and §5.2 claim that augmenting each solver 'consistently improves' reconstruction quality. This is not supported by the reported metrics: in Table 1, Blind Deblur, BON has higher PSNR (25.15) than RFJS (25.04); in Table 2, Motion Deblur, BON has higher PSNR (19.24) than RFJS (18.61). Many PSNR/SSIM differences are ~0.1–0.3 dB, likely within sampling noise. The main tables lack error bars and significance tests; only Appendix B (Tables 3–4) includes standard deviations. Please add error bars and statistical significance (e.g., paired tests) across all main tables, and soften the 'consistently improves' claim to reflect metric-specific and task-specific gains.
  2. [Appendix C and Algorithm 1] The implemented Resample step is 'greedy resampling, i.e., we pick the best candidate within each group of size g_t and replicate it g_t times' (Appendix C), not the weighted resampling with Cat(w[G_i]) specified in Algorithm 1 and analyzed in Proposition 1. The theoretical convergence as N→∞ applies to weighted resampling; the greedy rule is a heuristic. The statement that greedy resampling has 'similar effect as using smaller B with moderate temperature' is not demonstrated. Please either implement the weighted resampling to match the theory, or add experiments/theory comparing greedy vs. weighted resampling, and clearly state that the theoretical guarantees do not cover the deployed algorithm.
  3. [§4.1, Eq. (2)] The modeling assumption p_{0|S}(x0|s) ∝ p0(x0) exp(r(x0;s)/τ) is load-bearing: the entire search is driven by the reward. The paper provides no calibration or negative-control experiments to show that the chosen rewards (AdaFace distance, ImageReward, NMI) are monotone surrogates for the true compatibility with side information. In severely ill-posed settings (×32 SR, 16× MRI undersampling), a miscalibrated reward could cause greedy selection of a reconstruction that matches s but not the ground truth. Please add mismatched-side-information experiments (e.g., wrong-identity reference, semantically inconsistent text, scrambled MRI contrast) and/or an analysis of reward alignment to demonstrate that maximizing the reward moves samples toward the true x0.
  4. [§5.2, Table 1 and FS metric] The largest reported gains are on FaceSimilarity (FS), which is computed using the same AdaFace embedding that defines the reward. Part of the FS improvement is therefore by construction. Independent grounding is provided by PSNR/SSIM/LPIPS, but those gains are small and sometimes reversed. Please report a metric that is not aligned with the reward, such as identity retrieval accuracy with a different face-recognition model, or analyze the correlation between the reward and ground-truth fidelity, to separate genuine reconstruction improvement from reward overfitting.
  5. [§5.2, Table 2 (CS definition)] The paper defines CLIPScore as the cosine similarity between CLIP embeddings of the ground truth and reconstruction, but the side-information reward is ImageReward, a text-image alignment model. It is unclear how well CS measures the effect of the text side information. Please also report the actual ImageReward score (the optimized reward) to verify that search increases the targeted reward, and clarify the relation between CS and the text condition.
minor comments (5)
  1. [Figure 6 / Table 6 reference] The text says 'Table 6 shows our methods consistently outperform' but the results are presented in Figure 6. Correct the cross-reference.
  2. [§5.2, MRI paragraph] The MRI paragraph says 'We pair PD and PDFS contrasts' but Figure 5 captions and Figure 6 are not fully self-contained. Consider clarifying the reconstruction target and which contrast is side information.
  3. [Equation (6) and η] The derivation of Eq. (6) introduces η as a replacement for 1/(2σ_y^2). It would help to state explicitly that η is treated as a hyperparameter and to discuss its range/selection in Appendix C.
  4. [Section 4.1, Theorem 3 reference] The text refers to 'Theorem 3, deferred to Appendix A.2' but the theorem is numbered Proposition 3 in the appendix. Renumber or cross-reference consistently.
  5. [Typos throughout] Some typos and formatting issues appear: 'candiate' (§4.1), 'inverse problems' in the introduction used without hyphens, and the captions of Figures 3–5 have inconsistent 'Side Measurement' headers. A careful proofread is recommended.

Circularity Check

1 steps flagged

One reported metric (NMI) is the same quantity being optimized; otherwise the derivation is an explicit modeling assumption and PSNR/SSIM/LPIPS provide independent evidence.

specific steps
  1. self definitional [Section 5.2, MRI with multi-contrast side information; Figure 6 table; Algorithm 1 line 8]
    "We use normalized mutual information (NMI) as reward, which is robust to contrast changes. Table 6 shows our methods consistently outperform the baseline in all the metrics of interest."

    The reported NMI column is the same NMI used as the side-information reward. Algorithm 1 resamples particles using r[i] = r(hat x_{0|t,Y}[i]; s) and selects the final sample 'via reward maximization.' Thus a higher NMI than a baseline that does not search is enforced by the selection rule, not an independent validation of reconstruction quality. This makes the NMI-specific improvement circular by construction; the PSNR, SSIM, and LPIPS gains are external to the optimized reward and retain independent evidentiary value.

full rationale

The main derivation is not circular. Eq. (2) is explicitly introduced as a modeling assumption ('we make the following modeling assumption'), and Proposition 1 is a valid consequence of that assumption plus Bayes' rule; no fitted parameter is later relabeled as a prediction. The paper does not rely on a self-citation chain: no cited prior work is by the present authors, and the KL-regularized reward-tilting result is attributed to standard external work (Rafailov et al., Ouyang et al.). The central empirical claim that search improves reconstruction is supported by PSNR, SSIM, LPIPS, and CLIPScore, which are not identical to the optimized rewards. The only concrete objective/metric overlap is the MRI experiment, where NMI is both the reward and a reported evaluation metric, making that specific improvement partial-by-construction. FaceSimilarity is closely aligned with the AdaFace-based reward, but it is evaluated against the ground truth rather than the side-information image, so it is not a strict identity. The greedy-resampling implementation is a heuristic, but the paper explicitly concedes in Section D that it lacks formal exploration-exploitation guarantees, so this is a robustness concern rather than circularity.

Axiom & Free-Parameter Ledger

5 free parameters · 6 axioms · 0 invented entities

The central claim rests on a small set of tunable hyperparameters (B, N, η, τ, greedy-resampling rule) and on the assumption that pretrained reward models are faithful surrogates for side-information compatibility. No new physical entities are introduced; the reward-tilted posterior in Eq. (2) is a mathematical modeling device, not an independently evidenced object.

free parameters (5)
  • temperature τ = not reported
    Introduced in Eq. (2) to define the reward-tilted posterior and in the resampling weights exp(r/τ). No value is reported in experiments; the greedy resampling implementation makes its role unspecified.
  • resampling base B = 16 (DPS face), 4 (DAPS), 8/16 (MPGD), 100 (text), 512/128/32 (2D toy)
    Chosen per task and sampler in Appendix C. For MPGD Gaussian deblur, B=16 was chosen because B=8 'becomes detrimental', indicating tuning to maximize metrics.
  • particle count N = 8 (face/MRI), 4 (text)
    N is chosen per experiment; Appendix B.6 shows FaceSimilarity improves with N, so the absolute gains depend on this compute budget.
  • gradient scale η = 0.5 (DPS/MPGD box), 0.25 (MPGD SR/Gaussian), 13/15 (DAPS)
    Used in Eq. (6) and Algorithm 1 to compute the reward input. Appendix B.5 shows high scales degrade metrics, so values are tuned per task and sampler.
  • greedy resampling rule = pick best candidate in each group
    Appendix C replaces the stochastic Cat(w) resampling with deterministic best-in-group selection. This is a modeling choice not accounted for in the theoretical particle-filter derivation.
axioms (6)
  • ad hoc to paper p_{0|S}(x0|s) ∝ p0(x0) exp(r(x0;s)/τ) (Eq. 2)
    Stated as a modeling assumption in Sec. 4.1 and used to derive Proposition 1 and Eq. (5). No empirical validation that real side-information posteriors follow this exponential-tilt form.
  • domain assumption The reward function r orders reconstructions consistently with p_{0|S}
    Sec. 4.1 assumes if r(x1,s)>r(x2,s), then x1 is more compatible with s. Applied to AdaFace, ImageReward, and NMI without testing against mismatched or misleading side information.
  • domain assumption DPS-style approximation pY|t(y|xt) ≈ pY|0(y|hat x0|t)
    Inherited from Chung et al. and used in Lemma 2, Proposition 3, and Algorithm 1. This is an approximation, not an exact identity.
  • standard math Markov/conditional independence: pS|0,t,t+1,Y = pS|0
    Used in the proof of Proposition 1 in Appendix A.1. Reasonable but assumes side information depends on x0 only through the clean signal, not the noisy diffusion state.
  • domain assumption Finite-N particle resampling approximates the target posterior
    Sec. 4.2 acknowledges convergence only as N→∞. Experiments use N=4 or 8 with greedy resampling, so the posterior interpretation is heuristic.
  • domain assumption Pretrained unconditional diffusion prior matches the target data domain
    FFHQ prior used on Celeb-HQ faces, ImageNet prior on ImageNet images. Assumes no significant domain gap affecting the reward-guided search.

reviewed 2026-08-04 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Inference-Time Search Using Side Information for Diffusion-Based Image Reconstruction." pith.science (2026). https://pith.science/paper/55X6Q3SY

@misc{pith2026251003352,
  author       = {Pith},
  title        = {Pith review of: Inference-Time Search Using Side Information for Diffusion-Based Image Reconstruction},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/55X6Q3SY}},
  note         = {Machine review of arXiv:2510.03352}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Diffusion models have been used as priors for solving inverse problems. However, existing approaches typically overlook side information that could significantly improve reconstruction quality, especially in severely ill-posed settings. In this work, we propose a novel framework that incorporates side information into existing diffusion-based inverse problem solvers via inference-time search, in a plug-and-play, training-free manner. Through extensive experiments across a range of inverse problems, including inpainting, super-resolution, and several deblurring tasks, and across multiple diffusion-based inverse problem solvers (DPS, DAPS, and MPGD), we show that augmenting each solver with our framework consistently improves the quality of the reconstructions over the corresponding original method. To demonstrate the generality of our approach, we consider diverse forms of side information, including reference images, textual descriptions, and anatomical MRI scans. The code is available at this \href{https://github.com/mahdi-farahbakhsh/DISS}{repository}\footnote{https://github.com/mahdi-farahbakhsh/DISS}.

Figures

Figures reproduced from arXiv: 2510.03352 by Dileep Kalathil, Jean-Francois Chamberland, Krishna Narayanan, Mahdi Farahbakhsh, Vishnu Teja Kunde.

Figure 1
Figure 1. Figure 1: Illustration of the performance of our inference-time search algorithm for using side information [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Illustration of the group size resampling strategies of different search algorithms. [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Image as side information: Qualitative illustration of the performance of our RFJS algorithm compared to the DPS baseline on linear and nonlinear inverse problems. RFJS is able to capture many details that are missed by the DPS baseline to achieve a superior reconstruction quality. N/2 that are resampled independently; at every B/4 steps, groups of size N/4 are resampled; and so on. This hierarchical sched… view at source ↗
Figure 4
Figure 4. Figure 4: Text as side information : Qualitative illustration of the performance of our RFJS algorithm compared to the DPS baseline. For example, the side information provided for the super resolution task is ‘golden retriever sitting on a snowy frozen lake, facing forward’. RFJS is able to capture many details that are missed by the DPS baseline to achieve a superior reconstruction quality. MPGD are deferred to App… view at source ↗
Figure 5
Figure 5. Figure 5: Contrast Image as Side Information: Qualitative MRI reconstruction with RFJS vs. ContextMRI. The shapes and line edges are well preserved in our reconstruction. PDFS with PD Algorithm PSNR (↑) SSIM (↑) LPIPS (↓) NMI (↑) RFJS 25.85 0.801 0.375 0.457 GS 25.33 0.797 0.375 0.455 BON 25.47 0.797 0.376 0.454 ContextMRI 25.39 0.795 0.383 0.451 PD with PDFS RFJS 27.85 0.920 0.358 0.579 GS 27.80 0.920 0.360 0.579 B… view at source ↗
Figure 7
Figure 7. Figure 7: Qualitative illustration of the relevance of the FaceSimilarity metric and the superior performance [PITH_FULL_IMAGE:figures/full_fig_p012_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Additional samples using DPS as the base sampler. [PITH_FULL_IMAGE:figures/full_fig_p023_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Qualitative comparison on ImageNet with textual side information. For highly degraded obser [PITH_FULL_IMAGE:figures/full_fig_p023_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Qualitative comparison of algorithms using DAPS as the base sampler. Our method offers better [PITH_FULL_IMAGE:figures/full_fig_p024_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Qualitative comparison of algorithms using MPGD as the base sampler. Our method offers [PITH_FULL_IMAGE:figures/full_fig_p025_11.png] view at source ↗
Figure 13
Figure 13. Figure 13: Illustration of the effect of B in utilization of side information for the reconstruction in a linear inverse problem with a mixture of Gaussian prior. of the clean signal from the noisy state xt . Using this reward, RFJ Search produces posterior samples p(x0 | y, s) as shown in the three rightmost images of [PITH_FULL_IMAGE:figures/full_fig_p026_13.png] view at source ↗
Figure 12
Figure 12. Figure 12: Comparison of perfor￾mance of RFJ Search (RFJS) and Greedy Search (GS) as a function of B for a randomly generated reward net￾work rθ. RFJS outperforms GS across all values of B. To compare RFJ and Greedy Search, we next consider a more realistic scenario where side information is generated through a neural network reward model. For example, in face reconstruction tasks, side information may come from the… view at source ↗
Figure 14
Figure 14. Figure 14: Effect of reward-gradient guidance in diffusion-based inverse problems. We show 4 runs with different random seeds (rows), and for each seed we vary the gradient scale across 4 settings (columns). Within each row, the noise realization is identical and only the gradient scale changes; within each column, the gradient scale is fixed while the random seed varies. The ground truth and degraded input are the … view at source ↗
Figure 15
Figure 15. Figure 15: Qualitative comparison of the effect of gradient scale on reconstruction paths. [PITH_FULL_IMAGE:figures/full_fig_p028_15.png] view at source ↗
Figure 16
Figure 16. Figure 16: Qualitative comparison of the effect of gradient scale on reconstruction paths. Notice that while [PITH_FULL_IMAGE:figures/full_fig_p028_16.png] view at source ↗
Figure 17
Figure 17. Figure 17: Qualitative comparison of the effect of gradient scale on reconstruction paths for super resolution [PITH_FULL_IMAGE:figures/full_fig_p028_17.png] view at source ↗
Figure 18
Figure 18. Figure 18: Scaling of search algorithms with respect to the number of particles. [PITH_FULL_IMAGE:figures/full_fig_p029_18.png] view at source ↗
Figure 19
Figure 19. Figure 19: Effect of the number of particles N on the FaceSimilarity metric in DPS. As N increases, we proportionally increase the base to encourage greater exploration. Specifically, for N = 2, 4, 8, the corresponding base values are B = 4, 8, 16. Using many particles with a small value for base B can overly emphasize reward maximization, which may degrade other evaluation metrics. B.6 Effect of Number of Particles… view at source ↗
Figure 20
Figure 20. Figure 20: Effect of number of particles N on the FaceSimilarity metric. RFJ Search algorithm offers the best scaling performance, followed by Greedy Search algorithm. Finally, BestOfN performance improves, but only marginally. 64 indicates the size of the box for inpainting, 6 indicates the down-sampling factor in super-resolution, and 5 is the intensity of the Gaussian kernel in Gaussian deblur. The base resamplin… view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. Iterative Inference-time Scaling with Adaptive Frequency Steering for Image Super-Resolution

    cs.CV 2025-12 unverdicted novelty 7.0

    IAFS is a training-free iterative inference-time scaling framework that uses adaptive frequency-aware particle fusion to resolve the perception-fidelity conflict in diffusion super-resolution models, outperforming pri...

Reference graph

Works this paper leans on

63 extracted references · 9 linked inside Pith · cited by 1 Pith paper

  1. [1]

    Anderson, B. D. (1982). Reverse-time diffusion equation models. Stochastic Processes and their Applications , 12(3):313--326

    1982

  2. [2]

    Atalık, A., Chopra, S., and Sodickson, D. (2025). A Trust-Guided Approach to MR Image Reconstruction with Side Information . IEEE Transactions on Medical Imaging

    2025

  3. [3]

    D., Ho, J., Tarlow, D., and van den Berg, R

    Austin, J., Johnson, D. D., Ho, J., Tarlow, D., and van den Berg, R. (2021). Structured denoising diffusion models in discrete state-spaces. In Advances in Neural Information Processing Systems

    2021

  4. [4]

    Bansal, A., Chu, H.-M., Schwarzschild, A., Sengupta, R., Goldblum, M., Geiping, J., and Goldstein, T. (2024). Universal Guidance for Diffusion Models . In The Twelfth International Conference on Learning Representations

    2024

  5. [5]

    J., Corff, S

    Cardoso, G., el idrissi, Y. J., Corff, S. L., and Moulines, E. (2024). Monte Carlo guided Denoising Diffusion models for Bayesian linear inverse problems. In The Twelfth International Conference on Learning Representations

    2024

  6. [6]

    Y., Fessler, J

    Chun, S. Y., Fessler, J. A., and Dewaraja, Y. K. (2012). Non-local means methods using ct side information for i-131 spect image reconstruction. In 2012 IEEE Nuclear Science Symposium and Medical Imaging Conference Record (NSS/MIC)

    2012

  7. [7]

    Chung, H., Kim, J., Kim, S., and Ye, J. C. (2023a). Parallel Diffusion Models of Operator and Image for Blind Inverse Problems . In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

    2023

  8. [8]

    T., Klasky, M

    Chung, H., Kim, J., Mccann, M. T., Klasky, M. L., and Ye, J. C. (2023b). Diffusion posterior sampling for general noisy inverse problems. In The Eleventh International Conference on Learning Representations

  9. [9]

    L., and Ye, J

    Chung, H., Lee, D., Wu, Z., Kim, B.-H., Bouman, K. L., and Ye, J. C. (2025). Contextmri: Enhancing compressed sensing mri through metadata conditioning. arXiv preprint arXiv:2501.04284

    Pith/arXiv arXiv 2025

  10. [10]

    Chung, H., Sim, B., Ryu, D., and Ye, J. C. (2022). Improving Diffusion Models for Inverse Problems using Manifold Constraints . In Advances in Neural Information Processing Systems

    2022

  11. [11]

    C., Milanfar, P., and Delbracio, M

    Chung, H., Ye, J. C., Milanfar, P., and Delbracio, M. (2024). Prompt-tuning latent diffusion models for inverse problems. In Proceedings of the 41st International Conference on Machine Learning

    2024

  12. [12]

    C., Milanfar, P., Dimakis, A

    Daras, G., Chung, H., Lai, C.-H., Mitsufuji, Y., Ye, J. C., Milanfar, P., Dimakis, A. G., and Delbracio, M. (2024). A survey on diffusion models for inverse problems. arXiv preprint arXiv:2410.00083

    Pith/arXiv arXiv 2024

  13. [13]

    and Nichol, A

    Dhariwal, P. and Nichol, A. (2021). Diffusion models beat gans on image synthesis. In Advances in Neural Information Processing Systems

    2021

  14. [14]

    and Song, Y

    Dou, Z. and Song, Y. (2024). Diffusion posterior sampling for linear inverse problem solving: A filtering perspective. In The Twelfth International Conference on Learning Representations

    2024

  15. [15]

    Efimov, T., Dong, H., Shah, M., Simmons, J., Donegan, S., and Chi, Y. (2025). Leveraging multimodal diffusion models to accelerate imaging with side information. In IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)

    2025

  16. [16]

    Efron, B. (2011). Tweedie’s formula and selection bias. Journal of the American Statistical Association , 106(496):1602--1614

    2011

  17. [17]

    J., Thielemans, K., Pizarro, L., Atkinson, D., Ourselin, S., Hutton, B

    Ehrhardt, M. J., Thielemans, K., Pizarro, L., Atkinson, D., Ourselin, S., Hutton, B. F., and Arridge, S. R. (2014). Joint reconstruction of pet-mri by exploiting structural similarity. Inverse Problems

    2014

  18. [18]

    J., Shlens, J., and Szegedy, C

    Goodfellow, I. J., Shlens, J., and Szegedy, C. (2015). Explaining and harnessing adversarial examples. In International Conference on Learning Representations

    2015

  19. [19]

    Z., Salakhutdinov, R., and Ermon, S

    He, Y., Murata, N., Lai, C.-H., Takida, Y., Uesaka, T., Kim, D., Liao, W.-H., Mitsufuji, Y., Kolter, J. Z., Salakhutdinov, R., and Ermon, S. (2024). Manifold preserving guided diffusion. In The Twelfth International Conference on Learning Representations

    2024

  20. [20]

    P., Poole, B., Norouzi, M., Fleet, D

    Ho, J., Chan, W., Saharia, C., Whang, J., Gao, R., Gritsenko, A., Kingma, D. P., Poole, B., Norouzi, M., Fleet, D. J., and Salimans, T. (2022). Imagen video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303

    Pith/arXiv arXiv 2022

  21. [21]

    Ho, J., Jain, A., and Abbeel, P. (2020). Denoising diffusion probabilistic models. In Advances in Neural Information Processing Systems

    2020

  22. [22]

    Hyder, R., Hegde, C., and Asif, M. S. (2019). Fourier phase retrieval with side information using generative prior. In 2019 53rd Asilomar Conference on Signals, Systems, and Computers

    2019

  23. [23]

    Jones, L. K. (2009). Local minimax learning of functions with best finite sample estimation error bounds: Applications to ridge and lasso regression, boosting, tree learning, kernel machines, and inverse problems. IEEE Transactions on Information Theory

    2009

  24. [24]

    Kawar, B., Elad, M., Ermon, S., and Song, J. (2022). Denoising diffusion restoration models. In Advances in Neural Information Processing Systems

    2022

  25. [25]

    Y., Chung, H., and Ye, J

    Kim, J., Park, G. Y., Chung, H., and Ye, J. C. (2025a). Regularization by texts for latent diffusion inverse solvers. In International Conference on Learning Representations

  26. [26]

    K., and Liu, X

    Kim, M., Jain, A. K., and Liu, X. (2022). Adaface: Quality adaptive margin for face recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    2022

  27. [27]

    Kim, S., Kim, M., and Park, D. (2025b). Test-time Alignment of Diffusion Models without Reward Over-optimization . In The Thirteenth International Conference on Learning Representations

  28. [28]

    and Szepesv\' a ri, C

    Kocsis, L. and Szepesv\' a ri, C. (2006). Bandit based monte-carlo planning. In Proceedings of the 17th European Conference on Machine Learning

    2006

  29. [29]

    Kong, Z., Ping, W., Huang, J., Zhao, K., and Catanzaro, B. (2021). Diffwave: A versatile diffusion model for audio synthesis. In International Conference on Learning Representations

    2021

  30. [30]

    Levac, B., Jalal, A., Ramchandran, K., and Tamir, J. I. (2023). MRI Reconstruction with Side Information using Diffusion Models . 2023 57th Asilomar Conference on Signals, Systems, and Computers

    2023

  31. [31]

    Li, X., Uehara, M., Su, X., Scalia, G., Biancalani, T., Regev, A., Levine, S., and Ji, S. (2025). Dynamic search for inference-time alignment in diffusion models. arXiv preprint arXiv:2503.02039

    Pith/arXiv arXiv 2025

  32. [32]

    Liu, R., Gao, J., Zhao, J., Zhang, K., Li, X., Qi, B., Ouyang, W., and Zhou, B. (2025). Can 1b llm surpass 405b llm? rethinking compute-optimal test-time scaling. arXiv preprint arXiv:2502.06703

    Pith/arXiv arXiv 2025

  33. [33]

    Mota, J. F. C., Deligiannis, N., and Rodrigues, M. R. D. (2017). Compressed sensing with prior information: Strategies, geometry, and bounds. IEEE Transactions on Information Theory

    2017

  34. [34]

    Na, D., Ji, S., and Kim, J. (2022). Unrestricted black-box adversarial attack using gan with limited queries. In European Conference on Computer Vision

    2022

  35. [35]

    F., Leike, J., and Lowe, R

    Ouyang, L., Wu, J., Jiang, X., Almeida, D., Wainwright, C., Mishkin, P., Zhang, C., Agarwal, S., Slama, K., Ray, A., Schulman, J., Hilton, J., Kelton, F., Miller, L., Simens, M., Askell, A., Welinder, P., Christiano, P. F., Leike, J., and Lowe, R. (2022). Training language models to follow instructions with human feedback. In Advances in Neural Informatio...

    2022

  36. [36]

    Oymak, S., Thrampoulidis, C., and Hassibi, B. (2013). Simple bounds for noisy linear inverse problems with exact side information. arXiv preprint arXiv:1312.0641

    Pith/arXiv arXiv 2013

  37. [37]

    W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al

    Radford, A., Kim, J. W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al. (2021). Learning transferable visual models from natural language supervision. In International conference on machine learning

    2021

  38. [38]

    D., Ermon, S., and Finn, C

    Rafailov, R., Sharma, A., Mitchell, E., Manning, C. D., Ermon, S., and Finn, C. (2023). Direct Preference Optimization: Your Language Model is Secretly a Reward Model . In Advances in Neural Information Processing Systems

    2023

  39. [39]

    Rombach, R., Blattmann, A., Lorenz, D., Esser, P., and Ommer, B. (2022). High-Resolution Image Synthesis with Latent Diffusion Models . In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

    2022

  40. [40]

    Rout, L., Chen, Y., Ruiz, N., Kumar, A., Caramanis, C., Shakkottai, S., and Chu, W.-S. (2025). RB -modulation: Training-free stylization using reference-based modulation. In The Thirteenth International Conference on Learning Representations

    2025

  41. [41]

    Rout, L., Raoof, N., Daras, G., Caramanis, C., Dimakis, A., and Shakkottai, S. (2023). Solving linear inverse problems provably via posterior sampling with latent diffusion models. In Advances in Neural Information Processing Systems

    2023

  42. [42]

    S., Arriola, M., Schiff, Y., Gokaslan, A., Marroquin, E., Chiu, J

    Sahoo, S. S., Arriola, M., Schiff, Y., Gokaslan, A., Marroquin, E., Chiu, J. T., Rush, A., and Kuleshov, V. (2024). Simple and effective masked diffusion language models. In Advances in Neural Information Processing Systems

    2024

  43. [43]

    Setlur, A., Nagpal, C., Fisch, A., Geng, X., Eisenstein, J., Agarwal, R., Agarwal, A., Berant, J., and Kumar, A. (2025). Rewarding progress: Scaling automated process verifiers for LLM reasoning. In The Thirteenth International Conference on Learning Representations

    2025

  44. [44]

    J., Guez, A., Sifre, L., Van Den Driessche, G., Schrittwieser, J., Ioffe, S., Green, A., Chen, X., and et al

    Silver, D., Huang, A., Maddison, C. J., Guez, A., Sifre, L., Van Den Driessche, G., Schrittwieser, J., Ioffe, S., Green, A., Chen, X., and et al. (2016). Mastering the game of go with deep neural networks and tree search. Nature

    2016

  45. [45]

    Singhal, R., Horvitz, Z., Teehan, R., Ren, M., Yu, Z., McKeown, K., and Ranganath, R. (2025). A general framework for inference-time scaling and steering of diffusion models. arXiv preprint arXiv:2501.06848

    Pith/arXiv arXiv 2025

  46. [46]

    V., Lee, J., Xu, K., and Kumar, A

    Snell, C. V., Lee, J., Xu, K., and Kumar, A. (2025). Scaling LLM test-time compute optimally can be more effective than scaling parameters for reasoning. In The Thirteenth International Conference on Learning Representations

    2025

  47. [47]

    Sohl-Dickstein, J., Weiss, E., Maheswaranathan, N., and Ganguli, S. (2015). Deep unsupervised learning using nonequilibrium thermodynamics. In Proceedings of the 32nd International Conference on Machine Learning

    2015

  48. [48]

    M., Zhang, Z., Hu, X., Qu, Q., and Shen, L

    Song, B., Kwon, S. M., Zhang, Z., Hu, X., Qu, Q., and Shen, L. (2024). Solving inverse problems with latent diffusion models via hard data consistency. In The Twelfth International Conference on Learning Representations

    2024

  49. [49]

    Song, J., Meng, C., and Ermon, S. (2021a). Denoising Diffusion Implicit Models . In International Conference on Learning Representations

  50. [50]

    Song, J., Vahdat, A., Mardani, M., and Kautz, J. (2023a). Pseudoinverse-Guided Diffusion Models for Inverse Problems . In International Conference on Learning Representations

  51. [51]

    Song, J., Zhang, Q., Yin, H., Mardani, M., Liu, M.-Y., Kautz, J., Chen, Y., and Vahdat, A. (2023b). Loss-Guided Diffusion Models for Plug-and-Play Controllable Generation . In Proceedings of the 40th International Conference on Machine Learning

  52. [52]

    and Ermon, S

    Song, Y. and Ermon, S. (2019). Generative modeling by estimating gradients of the data distribution. In Advances in Neural Information Processing Systems

    2019

  53. [53]

    and Ermon, S

    Song, Y. and Ermon, S. (2020). Improved techniques for training score-based generative models. In Advances in Neural Information Processing Systems

    2020

  54. [54]

    P., Kumar, A., Ermon, S., and Poole, B

    Song, Y., Sohl-Dickstein, J., Kingma, D. P., Kumar, A., Ermon, S., and Poole, B. (2021b). Score-based generative modeling through stochastic differential equations. In International Conference on Learning Representations

  55. [55]

    and Deligiannis, N

    Tsiligianni, E. and Deligiannis, N. (2019). Deep Coupled-Representation Learning for Sparse Linear Inverse Problems With Side Information . IEEE Signal Processing Letters

    2019

  56. [56]

    E., Yang, K

    Wu, K. E., Yang, K. K., van den Berg, R., Alamdari, S., Zou, J. Y., Lu, A. X., and Amini, A. P. (2024). Protein structure generation via folding diffusion. Nature communications

    2024

  57. [57]

    L., Naesseth, C

    Wu, L., Trippe, B. L., Naesseth, C. A., Cunningham, J. P., and Blei, D. (2023). Practical and asymptotically exact conditional sampling in diffusion models. In Advances in Neural Information Processing Systems

    2023

  58. [58]

    Xu, J., Liu, X., Wu, Y., Tong, Y., Li, Q., Ding, M., Tang, J., and Dong, Y. (2023). ImageReward: Learning and Evaluating Human Preferences for Text-to-Image Generation . In Advances in Neural Information Processing Systems

    2023

  59. [59]

    Ye, H., Lin, H., Han, J., Xu, M., Liu, S., Liang, Y., Ma, J., Zou, J., and Ermon, S. (2024). TFG: Unified Training-Free Guidance for Diffusion Models . In Advances in Neural Information Processing Systems

    2024

  60. [60]

    Yu, J., Wang, Y., Zhao, C., Ghanem, B., and Zhang, J. (2023). FreeDoM: Training-Free Energy-Guided Conditional Diffusion Model . In 2023 IEEE/CVF International Conference on Computer Vision (ICCV)

    2023

  61. [61]

    J., Defazio, A., Stern, R., Johnson, P., Bruno, M., et al

    Zbontar, J., Knoll, F., Sriram, A., Murrell, T., Huang, Z., Muckley, M. J., Defazio, A., Stern, R., Johnson, P., Bruno, M., et al. (2018). fastmri: An open dataset and benchmarks for accelerated mri. arXiv preprint arXiv:1811.08839

    Pith/arXiv arXiv 2018

  62. [62]

    Zhang, B., Chu, W., Berner, J., Meng, C., Anandkumar, A., and Song, Y. (2024). Improving diffusion inverse problem solving with decoupled noise annealing. arXiv preprint arXiv:2407.01521

    Pith/arXiv arXiv 2024

  63. [63]

    Zhang, K., Zhang, Z., Li, Z., and Qiao, Y. (2016). Joint face detection and alignment using multitask cascaded convolutional networks. IEEE Signal Processing Letters

    2016

This paper was first reviewed by deepseek-v4-flash on August 4, 2026.